akoha / django-lean

A framework for performing and analyzing split-test experiments in Django applications.

Clone this repository (size: 263.3 KB): HTTPS / SSH
$ hg clone http://bitbucket.org/akoha/django-lean

Welcome

django-lean aims to be a collection of tools for Lean Startups using the Django platform. Currently it provides a framework for implementing split-test experiments in JavaScript, Python, or Django template code along with administrative views for analyzing the results of those experiments.

In the spirit of lean, we have open-sourced this library as quickly as possible, meaning that it might not have all the bells and whistles one might hope for, in particular with regards to setup. Any competent Django/Python developers should have little trouble integrating django-lean with their own projects, but if you would like to see support for your favorite installation mechanism, consider contributing a patch.

Discussion Group

For discussions related to the use or development of django-lean please use our Google group.

Features

django-lean allows you to perform split-test experiments on your users. In brief, this involves exposing 50% of your users to one implementation and 50% to another, then comparing the performance of these two groups with regards to certain metrics.

Experiments Types

django-lean supports two kinds of experiments currently:

  • Anonymous Conversion experiments compare the achievement of goals you define (i.e. "register" or "add to cart") amongst two groups of anonymous users.
  • Registered Engagement experiments compare a quantitative measure of engagement that you define (i.e., activity, revenue, time on site, ...) amongst two groups of registered users.

There's no real reason why one couldn't measure engagement of anonymous users or conversions of registered users (i.e. "basic to pro") but we didn't need this, so they're not implemented (again, patches welcome!).

Experiment Reports

django-lean provides daily reports of experiment results, including confidence levels.

  • For conversion experiments, results and confidence are displayed per conversion goal type (and for 'any' goal). Confidence is calculated using the chi-square method.
  • For engagement experiments, confidence is calculated using the Student's t-test method.

Experiment reports are prepared using the update_experiment_reports management command. It's advisable to execute this command from a nightly cron-job.

Bot Exclusion

django-lean attempts to exclude non-human visitors from experiment reports by only recording data for visitors who have JavaScript enabled.

Experiment Management

Experiments may be defined, enabled, disabled, or promoted via the django-admin interface. You may also define experiments in your source tree and have them automatically loaded into the database (see experiments.loader.ExperimentLoader).

Each experiment has a state, which affects whether visitors are enrolled in the experiment, and whether they see the control or test case of the experiment.

  • disabled: No visitors are enrolled in the experiment. All visitors see the control case of the experiment, even if they were previously enrolled in the test group.
  • enabled: All visitors who encounter the experiment are enrolled randomly in either the test or control group, and see the corresponding case.
  • promoted: No visitors are enrolled in the experiment. All visitors see the test case of the experiment, even if they were previously enrolled in the control group.

New experiments start in the disabled state.

Experiment Implementation

django-lean makes it easy to implement experiments in Python, JavaScript, or Django templates. Here are some examples:

Python

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

from experiments.models import Experiment
from experiments.utils import WebUser

...

def my_view_func(request, *args, **kwargs):
    if Experiment.test("my_experiment_name", WebUser(request)):
        view = edit_profile_test
    else:
        view = edit_profile_control
    return view(request, *args, **kwargs)

Django Templates

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

{% load experiments %}

...

<p>If you like what you see,

{% experiment change_buy_to_italics control  %}
<b>buy now!</b>
{% endexperiment %}

{% experiment change_buy_to_italics test  %}
<i>buy now!</i>
{% endexperiment %}

</p>

JavaScript

(In your HTML template:)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

{% load experiments %}

<script>
<![CDATA[
{% include "experiments/include/experiments.js" %}
]]>
</script>

{% include "experiments/include/experiment_enrollment.html" %}


...

{% clientsideexperiment <experiment_name> %}

(In your JavaScript:)

1
2
3
4
5

if (experiment.test("<experiment_name>")) {
  // test case
} else {
  // control case
}

Conversion Tracking

Conversion experiments track the rate of conversion for their test and control groups. It is up to you to define and record the achievement of one or more project specific conversion goals.

Defining Conversion Goals

Conversion goals are defined by placing rows in the experiments_goaltypes table. This table is not currently exposed via django-admin but probably should be (patches welcome!). Alternative ways to populate it include manually via SQL, manually via the Django shell, via your initial_data fixture, or by defining a data migration in your database management tool (we use django-south).

Here is an example of defining a goal type using the Django shell:

1
2
3
4
5
6
7
8
9

erik-wrights-macbook-pro:akoha erikwright$ ./manage.py shell --plain
Python 2.6.2 (r262:71600, Jul 16 2009, 12:11:28) 
[GCC 4.0.1 (Apple Inc. build 5490)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
(InteractiveConsole)
>>> from experiments.models import GoalType
>>> GoalType.objects.create(name="signup")
<GoalType: <GoalType: signup>>
>>>

Recording Conversions

Once you have defined a conversion goal type, you must record it when it is achieved place. This may be done either programatically, or using a tracking pixel.

Recording Conversions Programatically
1
2
3
4
5
6
7
8
9

from experiments.models import GoalRecord
from experiments.utils import WebUser

...

def my_signup_view_func(request, *args, **kwargs):
    # ... process the signup request
    GoalRecord.record("signup", WebUser(request))
    # ...

Recording Conversions Using a Tracking Pixel

Sometimes a conversion happens somewhere that you don't control (for example, an e-commerce platform). In this case, you can record the conversion by placing a transparent 1x1 pixel on the page that users see after the conversion occurs (i.e., the 'Thank You' page after purchase).

1

<img src="http://example.com/experiments/goal/purchase" height="1" width="1" />

Engagement Tracking

Engagement experiments track an arbitrary engagement value for each user in their test and control groups. It is up to you to define a function that calculates an appropriate engagement value for your users.

Here is an example engagement calculator:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

class MyEngagementScoreCalculator(object):

    def calculate_user_engagement_score(self, user, start_date, end_date):
        """
        Define engagement as 'dollars spent per day'
        """

        days_in_period = (end_date - start_date).days + 1

        period_purchase_total = sum([p.subtotal for p in Purchase.objects.filter(
                                                        user=user, date__gte=start_date, date__lte=end_date)])
        engagement_score = ((float)(period_purchase_total) /
                                                        days_in_period)
        return engagement_score

Your engagement calculator must be registered in settings.py as follows:

1
2
3

...
LEAN_ENGAGEMENT_CALCULATOR = "mycompany.MyEngagementScoreCalculator"
...

Dependencies

django-lean has a number of dependencies:

Mox and Beautiful Soup are used exclusively by unit tests. JQuery is used only to execute a single, trivial AJAX request and could easily be removed from the dependency list if one were motivated (patch please!).

You may optionally use South in order to facilitate migrations of the django-lean database schema, but it is not required:

Installation

django-lean has been developed with Django 1.0. Unit Tests run successfully with Django 1.1 but it has not been tried in production. If you successfully run it with another version, please update this documentation.

  1. Install django-lean using easy_install
  2. Add experiments to INSTALLED_APPS in settings.py
  3. Ensure that django.core.context_processors.request is in TEMPLATE_CONTEXT_PROCESSORS in settings.py
  4. Run manage.py syncdb to set up the django-lean tables.
  5. Run manage.py test experiments to see if everything is set up correctly.
  6. For every page that will contain an experiment (or in the response after a server-side experiment):
    1. Ensure that JQuery is included.
    2. Ensure that experiments/include/experiments.js is somehow included (perhaps copy it where your static files go, include it as part of your existing generated JS files, map it from urls.py, include it directly in a <script/> tag, etc.).
    3. Ensure that experiments/include/experiment_enrollment.html is rendered by your template.
  7. Install the admin and public url mappings in your site urls.py
  8. Register your engagement calculator in settings.py.
  9. Define one or more conversion goal types.
  10. Add conversion goal recording where appropriate.
  11. Define, implement, and enable an experiment
  12. Call manage.py update_experiment_reports nightly.
  13. Experiment, learn, repeat!

Installing URL Mappings

The following snippet added to urls.py should properly install the needed URL mappings (adjust to meet your needs):

1
2
3
4
5
6

...
urlpatterns += patterns('',
    url(r'^admin/django-lean/', include('experiments.admin_urls')),
    url(r'^django-lean/', include('experiments.urls')),
)
...

Development

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65

erik-wrights-macbook-pro:~ erikwright$ hg clone http://bitbucket.org/akoha/django-lean/
destination directory: django-lean
requesting all changes
adding changesets
adding manifests
adding file changes
added 46 changesets with 155 changes to 69 files
updating working directory
65 files updated, 0 files merged, 0 files removed, 0 files unresolved
erik-wrights-macbook-pro:~ erikwright$ cd django-lean
erik-wrights-macbook-pro:django-lean erikwright$ /usr/bin/python bootstrap.py 
Creating directory '/Users/erikwright/django-lean/bin'.
Creating directory '/Users/erikwright/django-lean/parts'.
Creating directory '/Users/erikwright/django-lean/eggs'.
Creating directory '/Users/erikwright/django-lean/develop-eggs'.
Generated script '/Users/erikwright/django-lean/bin/buildout'.
erik-wrights-macbook-pro:django-lean erikwright$ ./bin/buildout 
Develop: '/Users/erikwright/django-lean/.'
package init file 'experiments/tests/data/__init__.py' not found (or not a regular file)
Getting distribution for 'djangorecipe'.
Got djangorecipe 0.20.
Getting distribution for 'zc.recipe.egg'.
Got zc.recipe.egg 1.2.2.
Installing django.
django: Downloading Django from: http://www.djangoproject.com/download/1.1.1/tarball/
Getting distribution for 'mox'.
zip_safe flag not set; analyzing archive contents...
Got mox 0.5.0.
Getting distribution for 'BeautifulSoup'.
Got BeautifulSoup 3.1.0.1.
Generated script '/Users/erikwright/django-lean/bin/django'.
Generated script '/Users/erikwright/django-lean/bin/test'.
erik-wrights-macbook-pro:django-lean erikwright$ ./bin/test
Creating test database...
Creating table auth_permission
Creating table auth_group
Creating table auth_user
Creating table auth_message
Creating table django_content_type
Creating table django_session
Creating table django_site
Creating table experiments_anonymousvisitor
Creating table experiments_goaltype
Creating table experiments_goalrecord
Creating table experiments_experiment
Creating table experiments_participant
Creating table experiments_dailyengagementreport
Creating table experiments_dailyconversionreport
Creating table experiments_dailyconversionreportgoaldata
Installing index for auth.Permission model
Installing index for auth.Message model
Installing index for experiments.AnonymousVisitor model
Installing index for experiments.GoalRecord model
Installing index for experiments.Experiment model
Installing index for experiments.Participant model
Installing index for experiments.DailyEngagementReport model
Installing index for experiments.DailyConversionReport model
Installing index for experiments.DailyConversionReportGoalData model
..............................
----------------------------------------------------------------------
Ran 30 tests in 24.305s

OK
Destroying test database...
erik-wrights-macbook-pro:django-lean erikwright$

Other Resources

The following links might be of interest to those wanting to learn more about Lean Startup, split testing, and related concepts.


This revision is from 2010-06-20 19:26